Learning Based Image Selection for 3D Reconstruction of Heritage Sites

In this paper, we propose learning based pipeline with image clustering and image selection methods for 3D reconstruction of heritage site using cleaned internet sourced images. Cleaned internet sourced images means the images that do not contain an image with text, blur, occlusion, and shadow. 3D reconstruction of heritage sites is one of the emerging topics and is gaining importance as efforts are made to digitally preserve the heritage sites. 3D reconstruction using internet-sourced images is challenging as they often contain thousands of images taken from the same viewpoint. We propose to use autoencoders to extract robust features from images to cluster similar parts of heritage sites. We propose to use the image selection algorithm to select images from each cluster with the removal of redundant images. We demonstrate the proposed pipeline using available 3D reconstruction pipeline for a variety of heritage sites which contain one cluster to eight clusters and obtain better visual 3D reconstruction.